Does Schema Markup Really Improve AI Citations? What the Evidence Shows

You shipped the schema. Or you're about to, because someone forwarded you a post promising that structured data is how brands get cited by ChatGPT.

Then leadership asked what it bought you, and you didn't have a clean answer.

Take a breath. That gap between "we shipped schema" and "we can prove it worked" is normal, and it is not a sign you did something wrong. The claim has been repeated so often that it stopped sounding like a claim and started sounding like a fact.

So let's check it properly. Does schema help AI citations? Two controlled studies have now tested that question head-on, and the answer is more useful than either the hype or the backlash.

Here's what you'll get: what those studies actually measured, what Google says on the record, why the mechanism behaves the way it does, and where schema still pays you back. You'll finish knowing how much to invest and what to promise, without overclaiming to anyone.

So, Does Schema Help AI Citations? The Short Answer

No, not on its own. Adding schema markup to a page does not, on average, increase how often ChatGPT, Google AI Overviews, or Google AI Mode cite it.

That is what the controlled evidence shows as of mid-2026. One study even measured a small decline in AI Overview citations after schema was added.

Read that again, then let your shoulders drop. This does not mean schema is useless, and it does not mean you wasted the last two sprints. It means one specific marketing pitch, "add schema and AI will cite you," is not supported by controlled evidence.

Schema does real work. It just does a different job than the one it has been sold for.

Your action for this section is small: stop treating schema as your AI citation strategy. Keep it as infrastructure. We'll get to what it genuinely earns you further down.

What the Schema Markup AI Search Evidence Actually Tested

Two studies matter here, and they test different things. One measures outcomes at scale. The other opens the hood.

The large matched study

The first tracked 1,885 web pages that added JSON-LD schema between August 2025 and March 2026. Each treated page was matched to three control URLs on different domains with similar citation counts beforehand, giving roughly 5,655 controls.

That matched design is the part that matters. It is what separates "pages with schema get cited more" from "adding schema caused more citations."

The results across a 30-day before-and-after window:

The authors summed it up plainly: adding schema produced no major uplift in citations on any platform.

What about that −4.6%? Be careful with it. It is real in the data, and it is almost certainly not evidence that schema hurts you. The likeliest explanations are noise, or the messy reality that teams rarely add schema in isolation. Schema often ships alongside a redesign or a content rewrite, and those changes carry their own effects.

The honest headline is the null result, not the negative one.

One caveat you should not skip: every treated page already had 100 or more AI Overview citations before the test. So this tells you what happens when you add schema to a page already inside the consideration set. It cannot tell you whether schema helps a never-cited page earn its first citation.

The retrieval experiment

The second study is smaller and, for your purposes, more revealing. Researchers built a single test page for a fictional product and placed prices in eight different formats: visible HTML, JavaScript-rendered content, JSON-LD only, JSON-LD via JavaScript, hidden microdata, visible microdata, hidden RDFa, and visible RDFa.

Then they asked each AI system the same question five to ten times: what products are on this page, and what are the prices?

The scores were humbling. Gemini did best and still only extracted prices from four of the eight placements. ChatGPT managed three. Claude got zero. Perplexity and Google AI Mode could not read the page at all until it was indexed, and even then managed one and two.

One finding cuts through everything else. When the content lived only inside a JSON-LD block, no tested system extracted it. Not one.

When the same content was visible in the rendered HTML, systems could often read it, whether or not it carried microdata or RDFa attributes.

Format mattered less than visibility. That single sentence explains most of what follows.

Treat this one as a mechanism probe, not a measurement. It is one page and one product, so the direction is trustworthy and the precise numbers are not.

Why the "3x More Likely to Have Schema" Stat Misleads You

You have probably seen the stat that pages cited by AI engines are roughly three times more likely to have JSON-LD than pages that are not cited.

That number is real. It is also not evidence that schema causes citations, and this is the trap that has cost a lot of teams a lot of quarters.

Think about which pages tend to have schema. They sit on larger, better-maintained sites. Sites with more backlinks, more content, more engineers, and more SEO discipline. Those same properties are what earn citations.

So the schema and the citations share a common cause. They are not links in a chain.

It works like the classic example: ice cream sales predict drowning deaths. Both are true, both rise together, and neither causes the other. Summer causes both.

The matched study is the first to isolate the schema variable from the page quality variable. Once you control for quality, the effect disappears.

There is a harder truth inside that. Adding schema to a thin, poorly-built page will not make it cite-worthy. The markup describes the page; it does not improve it. If the underlying content does not deserve a citation, no amount of JSON-LD will argue it into one.

Here's your action, and it is genuinely one line: stop using the 3x stat to justify budget. If someone on your team is building a business case on it, they are building on sand, and it is kinder to say so now.

The Structured Data ChatGPT Never Sees

Why would schema do so little? The mechanism explains the null result, and once you see it you cannot unsee it.

Most AI answer engines work by retrieval. At answer time, they pull documents from an index and summarize what they pulled. What they can quote is what was in the retrievable, rendered content.

JSON-LD lives inside a script block. To most retrievers, that block is opaque metadata, not page text. It is structured data ChatGPT can walk straight past, because from its perspective there is nothing there to read.

Microdata and RDFa behave differently. They wrap visible text in attribute syntax, so the text still renders on the page. That is exactly why the retrieval experiment saw those formats work while hidden JSON-LD did not.

There is a design reason for this. Schema.org markup was built as a vocabulary for machines that already know to look for it: search engines building entity graphs, merchant feeds, recipe aggregators. It was never designed as a knowledge source for language model retrieval.

What each engine appears to be doing

The picture varies by engine, and the differences are worth knowing before you argue with anyone about them.

ChatGPT. Indexes pages and retrieves at answer time. In the retrieval test it read visible rendered HTML and walked past JSON-LD blocks. There is no public statement that it uses schema.org markup as a retrieval signal.

Google AI Overviews and AI Mode. These are built on Google's existing crawl and index pipeline, which definitely does parse structured data. What is missing is any public confirmation that structured data is used for citation selection rather than entity resolution and rich results. The citation numbers are consistent with no meaningful effect.

Perplexity. Runs its own crawler and index. It appears to treat schema.org markup mainly as a way to categorize pages, and there is no public evidence that JSON-LD beats well-structured visible content for citation rates.

Gemini. The strongest reader in the test. It handled visible content and was the only system that read JavaScript-rendered content. It still ignored JSON-LD.

Claude. Its web search leans on third-party search infrastructure, and it extracted nothing in the test, in any format. That suggests heavy reliance on its search provider's snippets rather than on markup.

Notice the pattern. Across five engines, the schema markup AI search evidence points the same way: visible content gets read, hidden markup gets skipped, and no engine advertises schema as a citation lever.

So the practical rule writes itself. If a fact matters, put it in visible prose on the page. Not only in the markup.

If your best answer to a buyer's question exists only inside a script block, it does not exist. Fix that before you touch anything else.

Does Schema Affect AI Overviews? What Google Says on the Record

Not directly, based on what Google publishes.

Google's structured data documentation describes structured data as a standardized format for classifying page content and helping Google understand it. It does not promise that valid structured data will produce rich results, and it does not promise a ranking benefit.

Google's AI Overviews guidance does not list structured data as a direct input either. The public advice for showing up in AI features points somewhere more familiar: clear original helpful content, clean crawlability and indexing, and page-level authority.

Be precise about what this proves. It is not proof that Google ignores structured data inside its AI pipelines. It is proof that Google does not market structured data as an AI citation lever. Those are different statements, and the difference is where credibility lives.

The FAQ story makes this concrete. Google deprecated FAQ rich results in August 2023 and pulled the feature from general display, then retired the reporting and testing support in 2026. The markup is still valid and Google still parses it. The visible payoff for most publishers is gone.

That matters because "add FAQ schema for AI Overviews" remains a common vendor line. For most sites, FAQ schema stopped producing a search feature years ago, and no controlled evidence shows it moves AI Overview citations either.

So when you ask "does schema affect AI Overviews," the fair answer is: not in any way Google claims or any study has demonstrated.

While you're here, four claims are worth pushing back on the next time they land in your inbox:

• "Schema boosts AI citations by up to 40%." No controlled study supports a number like this. The controlled study shows essentially zero.
• "AI engines crawl JSON-LD first." The retrieval experiment found the opposite. Hidden JSON-LD was the one thing nothing could read.
• "Add FAQ schema for AI Overviews." The rich result has been deprecated for most publishers since 2023, and no controlled evidence ties FAQ markup to AI Overview citations.
• "Structured data is required for AI visibility." Stronger than the evidence supports.

You do not need to win these arguments. You just need to stop funding them.

Where Schema Still Earns Its Keep

Here's the good news, and it is bigger than it sounds. Schema is still worth doing in 2026. Just for the documented reasons.

• Rich results. Product, Recipe, Article, Organization, LocalBusiness, BreadcrumbList, VideoObject, and Event markup still produce enhanced listings: stars, images, prices, sitelinks.
• Entity association. Organization, Person, and LocalBusiness markup with sameAs links to authoritative profiles feeds Google's Knowledge Graph. This shapes brand panels and disambiguation.
• Merchant integrations. Product, Offer, and AggregateRating markup feeds Shopping and free listings.
• News eligibility. NewsArticle and Person markup supports Top Stories eligibility.
• Voice reads. Speakable markup flags sections for assistants. Adoption is limited, but it is real for publishers who invested.
• Indexing hints. Attributes like datePublished and lastUpdated can support freshness signals and crawl efficiency.

Every one of those is supported. None of them is "ChatGPT will cite you."

Three caveats keep this honest, and you should carry them into any room where this gets debated.

Index time versus retrieval time. Schema could plausibly help a page get into a retrieval index in the first place. The retrieval experiment did not test that, and a citation-rate study of already-cited pages would not catch it. Real mechanism, unproven.

Schema type matters. The large study pooled all types together. Entity-rich schemas like Organization with sameAs may behave differently from content-wrapping schemas like Article. No published controlled study has separated them.

Time horizon. Thirty days may miss slow-burn effects. If schema improves entity resolution or crawlability, that could surface over months.

Two studies is a start, not a verdict. Say that out loud when you present this.

Schema Markup ROI AEO Teams Can Actually Bank On

So where should your effort go? Here is the decision, made simple.

Do implement schema, for the documented reasons. Rich results, entity association, merchant integrations. Treat it as SEO hygiene and infrastructure, not as a growth lever.

Do not expect schema alone to drive AI citations. Put that energy into content quality, originality, structure, and being genuinely quotable on the questions your buyers ask.

Prioritize by remaining value. Product, Organization, Article, BreadcrumbList, LocalBusiness, VideoObject, Event. Deprioritize FAQPage for general sites.

Validate before you deploy. Broken or spammy schema can trigger manual actions. That is a real downside risk on the other side of a null upside.

Match markup to visible content. Google explicitly warns against marking up content a user cannot see. Schema that misrepresents the page is a policy violation, not a clever shortcut.

Measure the right KPIs. Rich result impressions, click-through lift, entity panel appearance. Not AI citation rates, because there is no controlled evidence to act on there yet.

That last point deserves a beat, because it is where the honest schema markup ROI AEO conversation usually stalls. You still need to know whether AI engines mention and cite you at all. That question is worth answering directly, by tracking the prompts your buyers actually type and watching mention and citation rates move over time. Platforms like DeepSmith track exactly that across engines including ChatGPT, Gemini, Perplexity, Claude, and Google AI Mode, and they build schema, headings, and internal links into the writing pipeline rather than bolting them on later. Useful, and worth being clear about the limits: tracking tools are observational. They show you what is happening, not what caused it. No tool controls or guarantees citations, and any vendor promising otherwise is selling you the same unsupported claim this article just took apart.

Treat those dashboards as directional. Treat controlled studies as evidence. Keep the two straight and you will make better calls than most of the market.

If you want to see where you actually stand in AI answers before you spend another sprint on markup, start a free trial and look at your real numbers.

You are closer than you think here. You do not need to rip out your schema, and you do not need a new framework. You need to move it from the "growth strategy" column to the "hygiene" column, and put the freed-up hours into content worth citing.

That's one decision. Make it this week.